It's not generally made public as to the percentage of processors off a wafer which are functional. Both AMD and Intel release information on the number of processors/wafer, as well as occasionally releasing information on manufacturing costs, but as to the actual number of processors which are functional/wafer, that is not public. Some companies like TSMC, which manufactures chips for other companies, have information on process parameters and yields available to designers. AMD is actually keeping information on their yields very quiet right now. My "no more than 50%" figure above is based on talking with one of the professors here at OSU , who keeps very close tabs on semiconductor manufacturing technologies, and what he thought might be going on. AMD's former CEO was recently quoted denying low yields, but there's quite a buzz about it which seem to confirm that they are not good right now. To AMD's credit, they may be better now, but at least the first batch of Throughbred Athlons which are available now obviously were having some problems with exceeding 1800 MHz (2200+ speed rating). In general, the practical maximum speed at which a transistor can switch increases fairly linearly with a decrease in its dimensions. That's because by far the biggest problem right now in increasing clock speeds is the wires themselves which connect the transistors together.
Technical description (in other words I hope this helps you
):
The rate at which the voltage can change on a "wire" is going to be related to the supply voltage, capacitance, and resistance as: V(t)=Vinitial*(1 - exp(-t/RC)) or Vinitial*exp(-t/RC), depending on whether you are trying to pull it up (logic '1') or down (logic '0') respectively. The delays caused by the transistors themselves has become less important than the interconnect delay in the past several years. Capacitance (C) is: Koxide*Eo*Area/thickness of oxide, where Koxide is the dielectric constant of the insulating material used, which for SiO2 (which is the most common) is 3.9. Eo is the permitivity (sp) constant of free-space and is 8.85x10^-14 Farads/cm^2. Area is just the area of the wire, and the thickness is self explanatory. Resistance is related to the resistivity of the material used for the wires themselves, which for modern processes is going to be copper (and for some parts tungsten and some titanium or gold alloys), multiplied by the length/width of the wire.
When you move to .13u vs .18u, for example, if you just shrink everything, but leave the actual layout itself the same, the capacitance and resistance shrinks in a fairly linear fashion. This *should* allow a fairly linear increase in maximum clockspeeds. (I'm neglecting some fairly important issues, mostly related to quantum effects (i.e. tunneling) to keep this simple.) Another benefit is in power consumption. Power goes as P=1/2*clock speed*core voltage^2*capacitance. Since capacitance is decreased, so does power consumption. The problem is that since the die got smaller you have less contact area for a heatsink. The best way to reduce power consumption is to reduce voltage, but that comes at a price in maximum clockspeed (see above, remember V*exp(-t/RC)).
If you look at the layout photos of the Throughbred vs the Palomino core Athlon XPs (I've seen them at Tom's Hardware and Anandtech) you can see that for the most part everything just got smaller. The only real change is that the L2 cache was moved, but it's actually in a better location now. The fact that AMD hasn't been able to increase clockspeeds indicates that something else is at fault. Manufacturing (or a serious fault in the new layout) is going to be at fault here.
In AMD's case, note that they've only managed to reduce core voltage by only .05 to .1 Volts to maintain the same clockspeed. This is bad news, as it means power consumption is only going to be a little bit lower, and the real problem is that since Throughbred has a smaller die than Palomino you have less surface area to remove that heat. Once again, this relates back to manufacturing.
My educated opinion is that heat is actually limiting Throughbred right now, and that manufacturing issues are preventing the core voltage from being reduced, thus limiting the maximum clockspeed once again... (It's a never ending loop.
) Heat REALLY hurts how quickly transistors can switch. Heat disipation figures at Tom's Hardware confirm that the XP 2200+ draws only a couple watts less power than the XP 2100+. The smaller die size means that actual heat/die area is going to be much higher. Some people have actually managed to take an XP 2200+ to over 2500 MHz using water cooling, confirming what I've just said. Undoubtedly AMD will be able to get the clockspeeds up with a little bit of time and work in manufacturing T-bred, but what T-bred really needed was a heat-spreading metal plate like the P4 has. (Hammer does as well.)
Well I hope some of you were able to get something out of that at least...
